lywsvip/kwcode
1
0
Fork 0
mirror of https://github.com/val1813/kwcode.git synced 2026-09-03 06:34:30 +08:00
Code Issues Packages Projects Releases Wiki Activity
Files
797549450e0d41dabcdfdfefe14a95a922b84319
kwcode/STATUS.md
Val-sss fa1390f8a9 release: v1.6.1 版本号统一 + README/STATUS更新 + License Apache-2.0 
- pyproject.toml/formatters/telemetry/server/models/__init__ 全部统一到 1.6.1
- README: badge更新、更新日志加v1.6.0/v1.6.1条目
- STATUS: 版本号描述更新

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-05-07 09:23:19 +08:00

23 KiB
Raw Blame History

KWCode Project Status

Path: D:\program\codeagent2604\kwcode GitHub: https://github.com/val1813/kwcode Started: 2026-04-26 Goal: Local-model coding agent — maximize task completion rate via deterministic expert pipeline

Current: v1.6.1 (2026-05-07)

513 tests green + 67个bench tasks + 62个专项诊断测试。 架构收敛：删除WholeFileImplExpert/DependencyFixExpert，纯确定性机制驱动pipeline。 Generator增强：upstream_constraints注入system prompt + retry_hint携带上次代码 + tier=small填空框架。

v1.6.1 — Architecture Convergence (ExpertDirective收敛)

核心变更：删除独立Expert类，统一走pipeline

  • 删除 experts/whole_file_impl.py（238行）和 experts/dependency_fix.py（99行）
  • 存根实现：Generator通过ctx.gap自动解除函数限制（scope=whole_file），不需要独立Expert
  • 依赖安装：EnvProber在Phase0处理，不需要独立Expert
  • _select_moe_expert()始终返回None，所有任务走统一pipeline
  • GAP_TO_EXPERT_TYPE映射：NOT_IMPLEMENTED/STUB_RETURNS_NONE/MISSING_DEP → locator_repair
  • Gate VALID_EXPERT_TYPES移除whole_file_impl和dependency_fix

Generator增强

  1. _build_system()注入upstream_constraints到system prompt（之前只在prompt层注入，system层缺失）
  2. _build_retry_hint()携带上次生成的代码前300字符，LLM能看到自己的错误避免重复
  3. tier=small分支增加填空式编写规范（函数签名不变/只替换函数体/每个TODO 3-5行）

测试同步更新

  • test_stub_ratio_threshold.py：移除WholeFileImplExpert.can_handle()测试，改为验证GAP_TO_EXPERT_TYPE映射
  • test_routing_layer_stats.py：expected_expert从whole_file_impl/dependency_fix改为locator_repair

设计意图：bench结果可直接归因于GapDetector+scope+ExecutionStateTracker+EnvProber这套纯确定性机制，不被枚举专家类干扰判断。

v1.6.0 — MoE Deterministic Architecture

核心原则：LLM只做代码生成，所有路由/决策/状态判断全部确定性化。

GapDetector (core/gap_detector.py)

  • GapType enum（11种）：NONE/NOT_IMPLEMENTED/STUB_RETURNS_NONE/LOGIC_ERROR/MISSING_DEP/SYNTAX_STRUCTURAL/MISSING_TOOLCHAIN/WRONG_FILE/NO_TEST/ENVIRONMENT/UNKNOWN
  • Gap dataclass：gap_type + confidence + files + functions + error_msg + suggestion
  • GapDetector.compute()：纯正则匹配，零LLM调用，按优先级分类
  • GAP_TO_EXPERT_TYPE：确定性映射 GapType → expert_type（v1.6.1统一为locator_repair）

ExecutionStateTracker (core/execution_state.py)

  • TestDelta dataclass：每次修改后的测试状态变化
  • set_baseline() → record() → has_regression() → get_best_partial_state()
  • Git bisect式定位：知道哪步引入了问题，不盲目reset
  • 代码状态回滚交给Checkpoint，本类只追踪测试状态

EnvProber (core/env_prober.py)

  • 任务开始前确定性探测并修复环境（工具链+依赖+rig.json预构建）
  • LANG_TOOLCHAIN dict：6种语言的check/install/dep_cmd/dep_file
  • 缓存.kaiwu/env_profile.json（24h TTL，只缓存成功）
  • _find_working_test_cmd()：go用build验证（spec v2修正）

WholeFileImplExpert — 已删除（v1.6.1），功能由Generator通过ctx.gap scope=whole_file处理

DependencyFixExpert — 已删除（v1.6.1），功能由EnvProber在Phase0处理

Gate重构 (core/gate.py)

  • 确定性优先路由，LLM只做最后兜底二分类
  • 优先级：特殊任务关键词 → Gap路由(conf>=0.7) → 关键词匹配 → LLM兜底
  • _resolve_intent_vs_gap()：三层置信度消解（>=0.85 gap wins / 0.5-0.85 intersection / <0.5 user wins）
  • routing_source字段：记录每次路由决策来源

Orchestrator重构 (core/orchestrator.py)

  • Phase 0：EnvProber.probe_and_fix()（确定性环境修复）
  • Phase 1：无条件pre_test → GapDetector → ExecutionStateTracker.set_baseline()
  • Phase 2：Gap驱动expert_type覆盖（确定性优先于LLM分类）
  • Retry loop增强：回归检测→checkpoint.restore() / env_changed→_recompute_gap()
  • _select_moe_expert()：v1.6.1起始终返回None（统一走pipeline）

Verifier增强 (experts/verifier.py)

  • whole_file write_mode支持（直接写入整个文件）
  • _detect_wrong_file()：确定性检测修改文件与报错文件不匹配

审计日志增强 (audit/logger.py)

  • 目录分离：成功→/.kaiwu/logs/success/ / 失败→/.kaiwu/logs/failed/
  • 新增字段：routing_source, initial_gap_type, iterations[]
  • log_iteration()：每轮retry记录gap_type/expert_selected/can_handle_results/transition_reason/test_delta
  • 向后兼容：list_logs()/show_log()同时扫描新旧目录

TestParser (core/test_parser.py)

  • extract_failing_tests() / extract_passing_tests()
  • 纯正则，支持pytest/go/jest/rust四种格式
  • 供GapDetector、ExecutionStateTracker、Orchestrator共用

TaskContext新增字段 (core/context.py)

  • gap: Gap dataclass instance
  • confirmed_test_cmd: EnvProber提供的已验证测试命令
  • routing_source: 路由来源审计字段

专项诊断测试 (tests/diagnostic/, 62个测试)

  • test_gap_detector_accuracy.py：20个手工样本，GapType分类准确率>90%
  • test_stub_ratio_threshold.py：10+10文件，stub_ratio阈值验证>85% + GAP映射验证
  • test_execution_tracker_value.py：5个多迭代场景，回归检测+最优中间状态
  • test_routing_layer_stats.py：三层消噪触发率验证，gap_detector>llm_fallback

三飞轮系统（全部本地）

  • flywheel/strategy_stats.py — 错误策略有效性统计（~/.kwcode/strategy_stats.json）
    • record(error_type, sequence, success, retries) → 按error_type×sequence累计成功率
    • get_best_sequence() → min_attempts≥10时返回最优策略，否则用默认
    • 集成到orchestrator._record_flywheel()，每次任务完成后自动记录
  • flywheel/user_pattern_memory.py — 跨项目用户错误模式（~/.kaiwu/user_patterns.json）
    • record_task() → 统计error_type频率+成功率（滑动平均）
    • get_warning_hint() → 20+任务后生成中文提示注入ctx.kaiwu_memory
    • 5种错误类型各有针对性提示（syntax/assertion/import/runtime/patch_apply）
  • flywheel/skill_drafter.py — SKILL.md自动提炼（.kaiwu/skill_draft.md）
    • 30+成功轨迹后自动生成策略草稿
    • CLI: kwcode skill review/accept/discard

匿名遥测（opt-in，默认关闭）

  • telemetry/client.py — fire-and-forget daemon thread + httpx 3s超时
    • 只上传4字段：error_type, retry_count, success, model
    • 绝不上传：代码/路径/描述/用户身份
  • onboarding.py — init时询问opt-in（Confirm.ask，default=False）
  • CLI: kwcode telemetry status/enable/disable
  • config: telemetry_enabled 顶层字段，缺失=关闭（向后兼容）

服务端（已部署）

  • https://llmbbs.com → nginx反代 → 127.0.0.1:9753 (FastAPI+SQLite)
  • 3张表：task_events / daily_aggregates / strategy_effectiveness
  • API: POST /api/v1/event, GET /api/v1/health, GET /api/v1/stats
  • systemd: kwcode-telemetry.service, auto-restart
  • Let's Encrypt证书，自动续期

CLI增强

  • kwcode stats 增强：展示三飞轮状态+遥测状态
  • kwcode telemetry status/enable/disable — 遥测管理
  • kwcode skill review/accept/discard — SKILL.md草稿管理

集成点

  • orchestrator.init: +StrategyStats +UserPatternMemory +TelemetryClient
  • orchestrator.run(): ctx创建后注入user_pattern warnings + ctx._errors_encountered追踪
  • orchestrator._record_success/_record_failure_result: 调_record_flywheel()
  • _record_flywheel(): 策略统计 + 用户模式 + 遥测，三路全非阻塞

审计日志 (audit/logger.py)

  • AuditLogger: start() → log(stage, detail) → write(ctx, elapsed, success, model)
  • 存储：~/.kaiwu/logs/YYYY-MM-DD_HHMMSS_<expert_type>.json
  • 不记录代码内容，只记录：任务描述/Gate分类/专家执行时间/文件名/测试结果/重试次数
  • 最多保留100条，超出自动清理
  • orchestrator._emit()从@staticmethod改为实例方法，每个事件自动记录到audit
  • CLI: kwcode log / kwcode log show <id> / kwcode log clear

kwcode model命令 (cli/commands/model_cmd.py)

  • kwcode model — 显示当前模型配置+能力tier
  • kwcode model set <名称> — 切换模型（写入config.yaml）
  • kwcode model probe — 探测模型详情（Ollama API: family/参数量/量化/reasoning）

缩进对齐修复 (Generator._align_indentation)

  • 修复系统性bug：LLM返回class方法时丢失缩进（0空格 vs 原始4空格）
  • apply_patch替换后方法"跑出"class导致IndentationError
  • 修法：_generate_modified()返回后立刻调_align_indentation()补齐缩进差

P0: Hashline锚点编辑 (tools/hashline.py)

  • add_anchors(): 每行加6字符MD5哈希锚点 行号|哈希| 内容
  • parse_anchor_edits(): 解析 EDIT/DELETE/INSERT_AFTER 指令
  • apply_anchor_edits(): 验证哈希→应用编辑，任一哈希不匹配则拒绝全部
  • Generator首次尝试用HASHLINE_PROMPT（max_tokens=1024），失败fallback到完整函数生成
  • 效果：模型只输出编辑指令而非复现整个函数，减少输出token，消除patch_apply文本不匹配

P1: Think模式自适应 (core/think_config.py)

  • get_think_config(expert_type, difficulty) → {"think": bool, "budget": int}
  • 策略表：easy→think=off / medium→budget=512 / hard→budget=2048-4096
  • chat/office永远关闭think（不需要推理）
  • Generator根据think_config调整max_tokens：base + budget
  • orchestrator在Gate分类后自动设置ctx.think_config

P2: Fast/Slow双阶段推理（融入orchestrator retry loop）

  • 第一次尝试：fast think（默认think=off，快速生成）
  • 第一次失败：升级到slow think（budget=2048）
  • 第二次失败：最大think预算（budget=4096）
  • 与现有retry_strategy(0→1→2)正交，think_budget独立递增
  • 对reasoning模型(QwQ-32B等)效果最明显

测试：21个新测试（P0-P2）

  • Hashline: anchors/strip/parse/apply/mismatch/delete/insert/roundtrip (12)
  • ThinkConfig: easy/hard/chat/unknown/apply_tokens (8)
  • FastSlow: default/escalation (1)

Reviewer闭环（关键架构修复）

  • 原问题：Reviewer发现"改错文件"只记日志，不触发重试，36任务假成功
  • 修复：_record_success返回None → retry loop捕获 → 重置Generator/Verifier → 重试
  • review gap注入ctx.retry_hint，Generator下次修正方向
  • Reviewer prompt增强：注入initial_test_failure让LLM看到测试期望

Verifier 0/0假成功消除

  • tests_total=0时检查是否有测试文件存在
  • 有测试文件但0执行 → passed=False，报错"测试未执行"
  • 消除benchmark 36/37任务"audit成功但bench失败"的根因

Syntax熔断按tier区分

  • SMALL：1次重试后熔断（小模型重复同样错误）
  • MEDIUM/LARGE：2次重试后熔断（32B第一次syntax error常是偶然）

Test-First Loop（CC架构核心改动）

  • orchestrator.run()：locator_repair/refactor任务先调verifier.run_tests_only()拿测试报错
  • locator.locate_from_test_error()：从pytest/go test报错提取File+行号，精准定位（不靠语义搜索猜）
  • 优先级：test_error定位 > BM25+图 > LLM猜，test_error成功则跳过语义搜索
  • 解决benchmark36/37任务"audit成功但bench失败"的根因：Verifier 0/0假成功

Verifier修复（P0影响所有结果）

  • 测试文件发现：扫描project_root下*_test.py/test_*.py/*_test.go/*.test.ts，不只看tests/目录
  • 找不到测试文件才返回0/0，有测试文件但没在tests/时用文件路径直接跑pytest
  • 工具链检测：go/node/rust/java缺失时自动安装（apt-get），不再报syntax error
  • Go语法检查：go: not found不报语法错误（返回None跳过）

Locator增强

  • locate_from_test_error()：从Python/Go/TS测试报错提取文件名+行号+函数名
  • 过滤test文件和stdlib，只返回业务代码文件
  • import语句提取被测模块名，反向定位源文件

Prompt约束量化改造（CC风格）

  • GENERATOR_BASE_SYSTEM：定性→量化
    • "只做任务要求的事" → "每次patch只修改≤2个函数，修改行数≤30行"
    • "不要动无关代码" → "不触碰报错行±20行范围外的无关代码"
    • "不要加注释" → "不添加任何import/类型注解/docstring/注释到未修改的代码"
  • GENERATOR_PROMPT：
    • "只修改必要的部分" → "修改行数≤15行，不改动与错误无关的行"
  • RETRY_STRATEGIES hints：
    • syntax: "不改其他逻辑" → "修改≤5行，不触碰其他函数"
    • assertion: "只改最小代码" → "只改1个函数，修改≤10行"
    • unknown: "缩小修改范围" → "只修改1个函数，修改≤15行"
  • CHAT_SYSTEM：
    • "简短友好回复，2-3句话即可" → "≤100字回复，≤3句话"
  • CHAT_SEARCH_FAIL_SYSTEM：
    • 整段重写为"回复≤50字"硬约束

遥测防护（服务端三层守卫）

  • HMAC-SHA256签名：客户端用密钥签payload，无签名→403
  • IP限流：同一IP每分钟≤30次，超限→429
  • 字段校验：error_type枚举白名单 + model名正则([a-zA-Z0-9.:\-_/]) + 长度限制
  • /stats端点需管理token，无token→401

版本号统一

  • 唯一真相源：pyproject.toml version = "1.6.1"
  • formatters.py / telemetry/client.py / server/models.py / init.py 全部改为 importlib.metadata.version("kwcode")，未安装时 fallback "1.6.1"

测试：19个新测试

  • StrategyStats: record/get_best_sequence/min_attempts/persistence/corrupted recovery (5)
  • UserPatternMemory: record/warning_threshold/top_errors/summary/unknown_ignored (6)
  • SkillDrafter: draft_generation/save/exists/insufficient (4)
  • TelemetryClient: disabled_default/enabled_config/non_blocking/skip_disabled (4)

v1.5.0 — Isolated Search + Cross-File Contracts

Theory: WarpGrep (isolated search subagent) + CGM (graph-injected attention) + PENCIL (erase intermediate state) + SWE-ContextBench (context quality > model size)

SearchSubagent (experts/search_subagent.py)

  • Independent context window — search noise never enters Generator
  • Parallel file reads: ThreadPoolExecutor, 8 concurrent
  • Returns only precise {file, start_line, end_line, content}
  • Shadow TaskContext: Locator writes to shadow, main ctx stays clean

UpstreamManifest (core/upstream_manifest.py)

  • Deterministic AST extraction: Python ast module, regex fallback for others
  • Tracks: function signatures, constants, import dependencies
  • get_constraints_for_file() → injected into Generator prompt
  • check_consistency() → Verifier pre-check, catches arg count / constant mismatches
  • Zero LLM calls

PENCIL Compression + Contract Verification

  • task_compiler: _compact_subtask_result() keeps only signatures/constants/paths/test_status
  • orchestrator: locator step uses SearchSubagent, verifier pre-checks contracts
  • contract_violation error type triggers re-locate retry strategy
  • Generator prompt receives upstream_constraints + retry_hint

Code Quality (this release)

  • orchestrator.py run() split into 5 private methods (410→253 lines)
  • Full type annotations: Optional[DebugSubagent], Callable[[str,str],None]
  • all added to 7 core modules
  • pyproject.toml: license fixed, ruff + mypy configured
  • TUI: 30+ event icons added (contract_violation, ab_test, replay, etc.)
  • Server: /api/manifest endpoint, version 1.5.0

v1.4.0 — Multi-Language + TUI + IDE

Theory: XRAY MCP Server + OpenCode + CodeCompass

  • 7-language AST (Python/JS/TS/Go/Rust/Java/C#)
  • ast-grep with QUERY_TEMPLATES (LLM never writes patterns)
  • FastAPI server (port 7355) + SSE streaming
  • Textual TUI (file tree + event log + input)
  • VSCode extension (thin client, all logic server-side)

v1.3.0 — EventBus + Error Strategy + Cognitive Gate

Theory: Dive into Claude Code + Wink + ARCS + SpecEyes + OPENDEV + Turn-Control

  • EventBus: append-only log, replay, wildcard listeners
  • ToolGateway: per-expert permissions, file cache with dirty tracking
  • Error strategy routing: 6 error types → different retry sequences
  • CognitiveGate: diminishing returns detection → auto-stop
  • GraduatedCompactor: 3-layer progressive context compression
  • Plan auto-trigger for hard tasks
  • Worktree isolation (git worktree / tempdir fallback)
  • Speculative Prefetch: background file pre-read
  • SearchRouter: intent-aware routing (arXiv/S2/GitHub/PyPI/Open-Meteo)
  • Wink self-repair: scope_creep / repetitive_fix / patch_miss / empty_output

v1.2.0 — RIG Project Map

Theory: RIG + FastCode + CodeCompass

  • export_rig(): full project index (exports/imports/routes/test coverage)
  • upstream_summary structured dict for multi-task context passing
  • ConsistencyChecker: deterministic frontend/backend API mismatch detection
  • Gate/Locator prompt explicitly guided to query rig.json

v1.1.0 — P0+P1+P2 Optimizations

  • Verifier structured output (_classify_error: 5 error types)
  • Circuit breakers (syntax 1x, import immediate, same-type 3x streak)
  • Gate confidence scoring (0.92/0.75/0.55)
  • Experience Replay (BM25 similar trajectory lookup)
  • Session continuity (SessionState, 5-turn KWCODE.md re-injection)
  • Locator minimal context (function boundaries, 60-line cap, gap markers)
  • Watchdog 300s timeout
  • Gate accuracy stats (/stats command)

v0.9.0 — DAG Compiler + Debug Subagent

  • TaskCompiler: DAG scheduler, ThreadPoolExecutor + Kahn topological sort
  • Debug Subagent: sys.settrace variable capture on failure
  • Prompt Optimizer: trajectory → experience rules → YAML system_prompt
  • Cross-Encoder search reranking

Core Pipeline (v0.5.0–v0.8.0)

  • Gate → 6 pipeline routes (locator_repair/codegen/refactor/doc/office/chat)
  • BM25+AST call graph two-phase location (zero LLM, milliseconds)
  • Generator: original from file, LLM only generates modified
  • Verifier: syntax check + pytest
  • 3-stage retry + Reflection root cause analysis
  • 5 deterministic tools (read_file/write_file/run_bash/list_dir/git)
  • KWCODE.md project rules + /plan + Checkpoint + DocReader
  • Model capability tiers (SMALL/MEDIUM/LARGE)
  • Expert flywheel (trajectory → pattern → backtest → AB test → production)
  • 3-layer memory (PROJECT.md/EXPERT.md/PATTERN.md)
  • Office document generation (Excel/PPT/Word)
  • MCP Router, context compression, CJK BM25

Test Summary

Category Count Status
Core unit tests 38 PASS
Regression tests 173 PASS
P1 feature tests 33 PASS
P2 feature tests 21 PASS
Search refactor 19 PASS
Intent search 19 PASS
E2E real model 17 PASS
RIG modules 29 PASS
TaskCompiler 12 PASS
Multi-language 51 PASS
Server/TUI 16 PASS
SearchSubagent+Manifest 27 PASS
MoE Diagnostic 62 PASS
Total 513 All green

File Structure

kwcode/
├── pyproject.toml
├── README.md / README_zh.md
├── STATUS.md
└── kaiwu/
    ├── cli/
    │   ├── main.py              # 入口（173行）+ Typer路由
    │   ├── commands/task.py     # run/chat/vision/multi-task命令
    │   ├── commands/expert.py   # expert list/info/export/install
    │   ├── commands/config.py   # init/api/serve/setup-search
    │   ├── formatters.py        # Rich输出格式化
    │   ├── repl.py              # REPL交互循环
    │   ├── status_bar.py        # 状态栏(4档自适应)
    │   └── onboarding.py        # 首次启动引导
    ├── core/
    │   ├── event_bus.py         # Unified event bus (append-only + replay)
    │   ├── cognitive_gate.py    # Diminishing returns detection
    │   ├── wink.py              # Self-repair monitor
    │   ├── gate.py              # [v1.6] 确定性优先路由（Gap→关键词→LLM兜底）
    │   ├── orchestrator.py      # [v1.6] MoE pipeline + Gap驱动 + 回归检测
    │   ├── context.py           # TaskContext dataclass (+gap/confirmed_test_cmd/routing_source)
    │   ├── gap_detector.py      # [v1.6] GapType enum + GapDetector.compute() (zero LLM)
    │   ├── execution_state.py   # [v1.6] ExecutionStateTracker (regression detection)
    │   ├── env_prober.py        # [v1.6] EnvProber (toolchain/dep auto-fix, cached)
    │   ├── test_parser.py       # [v1.6] extract_failing/passing_tests (regex)
    │   ├── task_compiler.py     # DAG scheduler + WorktreeManager
    │   ├── upstream_manifest.py # [v1.5] Cross-file contract tracking (zero LLM)
    │   ├── planner.py           # /plan mode + risk assessment
    │   ├── checkpoint.py        # File snapshot (git stash / file copy)
    │   ├── kwcode_md.py         # KWCODE.md segmented loading
    │   ├── model_capability.py  # Model tier detection (SMALL/MEDIUM/LARGE)
    │   ├── context_pruner.py    # Context compression + GraduatedCompactor
    │   ├── network.py           # Network detection + proxy config
    │   └── sysinfo.py           # System info + VRAM monitoring
    ├── experts/
    │   ├── locator.py           # BM25+graph location + DocReader + Prefetch
    │   ├── search_subagent.py   # [v1.5] Isolated search (independent context)
    │   ├── generator.py         # [v1.6.1] Code generation + upstream_constraints system + retry last_code + small填空
    │   ├── verifier.py          # [v1.6] Syntax + pytest + whole_file + _detect_wrong_file
    │   ├── search_augmentor.py  # Search augmentation + BM25 rerank
    │   ├── consistency_checker.py # Frontend/backend API consistency (deterministic)
    │   ├── chat_expert.py       # Chat (search gating)
    │   └── office_handler.py    # Office document generation
    ├── search/                  # Intent-aware search routing
    ├── knowledge/               # PDF/Word/MD reader + CJK BM25
    ├── flywheel/                # Trajectory → pattern → generation → AB test
    ├── registry/                # Expert registry + .kwx packaging
    ├── notification/            # Flywheel notifications
    ├── stats/                   # Value tracking (SQLite)
    ├── memory/                  # 3-layer memory system
    ├── ast_engine/              # tree-sitter AST + call graph
    ├── server/                  # FastAPI + SSE (port 7355)
    ├── tui/                     # Textual TUI
    ├── mcp/                     # MCP Router
    ├── llm/                     # Ollama + llama.cpp backends
    ├── tools/                   # 5 deterministic tools + ToolGateway
    ├── audit/                   # [v1.6] Enhanced audit (success/failed split, iterations)
    └── tests/                   # 513 unit tests + 67 bench tasks + 62 diagnostic
        └── diagnostic/          # [v1.6] 4 architecture validation test suites

TODO

  1. CLI拆分：main.py 1861→173行 ✅ v1.5.0
  2. 注释统一中文 ✅ v1.5.0
  3. 专家细粒度EventBus emit ✅ v1.5.0
  4. bench tasks多语言覆盖（67题 Python/Go/TS） ✅ v1.5.0
  5. 删除WholeFileImplExpert/DependencyFixExpert，收敛到纯pipeline ✅ v1.6.1
  6. SQLite跨session查询
  7. pip publish到PyPI（v1.5.0）
  8. install.ps1 / install.sh一键安装
  9. SWE-bench评测（用评测VPS跑）
  10. 跑bench验证v1.6.1架构收敛效果

Known Issues

  • qwen3-vl:8b outputs in thinking field, content empty (thinking extraction added)
  • Reasoning models slow on Gate (8x multiplier)
  • SearXNG requires Docker Desktop, degrades to DDG without it
Reference in New Issue View Git Blame Copy Permalink